If AI coding is lowering your code quality, you're not managing quality right

TL;DR Highlight

Seven layers of defense cut bugs while output rose 2-3x with coding agents

Who Should Read

Backend or full-stack developers worried that adopting coding agents will raise their team's bug count

Core Mechanics

Have AI review requirements first
- After adopting spec-driven development, bugs in newly written code reportedly dropped sharply.
- Before that, up to a third of a feature's effort went to post-development bug fixing.
- The key step is having AI find gaps, edge cases, and conflicts with existing code in the requirements or tech design.
Tests first, implementation second
- Have the agent think through test scenarios from the requirements first.
- Then write tests, write the implementation, run the tests and fix issues, and backfill coverage gaps.
- This avoids tests that merely pass on bugs the agent just introduced.
Manual testing is not replaced
- Human testing of edge cases saw only modest productivity gains.
- That is why output rose 2-3x rather than 10x.
- E2E tests are only a rough check and don't replace manual testing.
Split quality checks into separate passes
- The author finds agents poor at following complex instructions in AGENTS.md or CLAUDE.md.

Community Reactions

Opinion that tests end up mirroring the implementation
- One commenter suggested running two sub-agents in parallel for tests and implementation so neither pollutes the other's context.
Skepticism about coverage numbers
- A commenter argued that 95%+ coverage says nothing about the solution itself.
- Another recommended a functional core, imperative shell design instead.
Experience applying it to legacy code
- One commenter used Claude on a 22-year-old legacy system for dependency updates, security audits, and new features.
- They raised coverage from 65% to 80%+ and still review code before merging.
Doubt about TDD with agents
- A commenter cited an HN article concluding that asking agents to do TDD wasn't particularly helpful.
- They worried about slower work, more tokens, and more code to review.
Limits of turning authors into editors
- One commenter said finding bugs in AI code can be harder than writing correct code yourself.

How to Apply

Run a requirements gap-review prompt
- Today, paste the spec of the feature you're building into an AI before implementing.
- Check the proposed edits yourself, since AI can invent issues.
Add quality passes to your implementation skill
- Split security, duplication, naming, and logic checks into separate passes in your agent workflow.
- Measure whether implementation time grows by 5-15 minutes.
Add two AI reviews plus a pruning pass to PRs
- Run reviews from different models on each PR.
- Add a final agent pass that filters out trivial comments.
Connect your error tracker to auto-diagnosis
- Have an agent root-cause errors captured by a service like Sentry and open fix PRs.

Try It

copy & run

# Paste into a Claude or ChatGPT chat (the article's Layer 1 method turned into a prompt)

Review the requirements document below. Do not write code.
Find: 1) missing requirements and edge cases 2) unexpected interactions with existing code 3) contradictions or ambiguities in the scenarios

[Requirements]
{{requirements document}}

[Existing code context]
{{relevant code or description}}

Give a one-line rationale for each problem, and mark anything that may not be a real issue as 'uncertain'.

Next step: add only the confirmed items to the requirements, then ask 'write test scenarios based on these requirements first'.

Terminology

spec-driven developmentWriting requirements and design docs before code and developing against them. Like drawing blueprints before building.

E2E testA test that runs a full user flow from start to finish. It's a test drive of the finished car, not its parts.

AGENTS.md / CLAUDE.mdInstruction files that tell a coding agent your project's rules. Think of an onboarding doc for a new hire.

MCPA connection standard that lets AI access external tools like browsers or logs. Think of it as a USB port for AI.